Workload Characterization of SPECweb2005

1
Workload Characterization of SPECweb2005

• Presentation at SPECworkshop
• by
• Rema Hariharan
• Ning Sun
• Sun Microsystems

2
SPECweb2005

• Banking workload Heavy dynamic content and
  fully secure
• Ecommerce workload Heavy dynamic content, some
  portion uses SSL
• Support workload Heavy static download, no SSL

3
SPECweb2005_Banking

• Users login, check account balance/transfer
  money/change profile, logout
• Includes GETs and POST
• Each dynamic request followed by request for
  image files (static)
• 304 responses
• 2 parallel TCP connections
• One SSL full handshake and one reuse

4
SPECweb2005_Banking Stats

• Average number of SW2005_Ops in a login session
  4.6
• Number of TCP connections made per login session
  2
• Number of Page requests per TCP connection 2.3
• Average number of HTTP requests per SW2005_Op
  12 (average derived from design parameters)
• So, the average number of HTTP requests per TCP
  connection 122.3 27.6
• Think Time between SW2005_Ops 9.98 sec
• Average number of SW2005 Ops/sec per user session
  0.149

5
SPECweb2005_Banking Network byte stats
6
Per 1000 user session statsBanking
7
SPECweb2005_Ecommerce

• Based on traffic seen at one of the major
  Ecommerce sites selling computers
• Uses both http and https requests
• Redirects and 304 responses emulated
• Heavy on back-end communication
• Includes searches, general browsing,
  customization and checkout (11 distinct scripts)
• Product images scale with the benchmark

8
Some stats for Ecommerce workload

• Average number of SW2005_Ops per incoming user
  session 8.8
• Number of http ops per SW2005_Op 17
• Number of SW2005_Ops/s per user session supported
  0.093
• Fraction of sessions that enter the check-out
  stage 2/3
• TCP connections without check out 2
• TCP connections with check-out 4
• TCP connections/sec per user session supported
  0.035

9
Network Characteristics Ecommerce
10
Per 1000 user session stats - Ecommerce
11
Other bottlenecks for ECommerce

• Memory more memory intensive than banking
• Performance difference between 8 GB memory
  configuration and 16 GB memory configuration can
  be substantial, 10-20
• BeSim Scaling

12
SPECweb2005_Support

• Based on patch download/support site traffic
  characteristics
• User logs in, searches for the patch and then
  downloads it.
• Includes dynamic pages and embedded images
• 304 responses
• Download size can be as large as 35 MB follows
  Zipf distribution for choice of directories and
  files.
• QOS defined separately for pages and downloads
• Designed to be intensive on Network and Disk IO.

13
Averages for Support workload

• Average number of SW2005_Ops in a login session
  14.5
• Number of TCP connections made per login session
  2
• Number of Page requests per TCP connection
  7.25
• Average number of HTTP requests per SW2005_Op
  21 (average derived from design parameters)
• So, the average number of HTTP requests per TCP
  connection 7.2521 152
• Think Time between SW2005_Ops 4.98 sec
• Average number of SW2005 Ops/sec per user session
  0.013

14
Summary Stats Support Workload
15
Network Characteristics Support workload
16
(No Transcript)
17
Zipf Parameter Analysis for Support workload
18
Questions to answer in the future

• Any other characteristic to represent?
• Do scripts represent real world scripts?
• How good is the representation of the composite
  score, currently a relative geometric mean?
• Is any characteristic represented incorrectly?
• Are the current workloads unfairly favoring any
  particular type of architecture, not representing
  reality?